Skip to content

perf: use fused gdn gating for qwen3.5 prefill. #1301

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
yingxudeng wants to merge 1 commit into from
Closed

perf: use fused gdn gating for qwen3.5 prefill. #1301

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 9 additions & 1 deletion xllm/compiler/tilelang/targets/ascend/kernels/fused_gdn_gating.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,15 @@
VECTOR_BYTES_PER_ITER = 256
SUPPORTED_NUM_HEADS = (4, 6, 8, 12, 16, 24, 32, 48, 64, 128)
MAX_VEC_CORE_NUM = detect_vec_core_num()
BATCH_SIZE_SPECIALIZATIONS = tuple(range(2, 49, 2))


def _build_batch_size_specializations() -> tuple[int, ...]:
small_dense = range(2, 49, 2)
large_sparse = range(64, DEFAULT_MAX_BATCH + 1, 64)
return tuple(sorted(set((*small_dense, *large_sparse))))


BATCH_SIZE_SPECIALIZATIONS = _build_batch_size_specializations()


def select_launch_block_num(*, num_batches: int, vec_core_num: int) -> int:
Expand Down
5 changes: 3 additions & 2 deletions xllm/core/kernels/npu/tilelang/fused_gdn_gating_wrapper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -193,8 +193,9 @@ void run_tilelang_fused_gdn_gating_chunk(const torch::Tensor& A_log,

auto specialization = build_runtime_specialization(a);
const auto* entry = find_fused_gdn_gating_kernel_entry(specialization);
// Expected fast path: compiled batch_size variants are dense [2, 4, ..., 48].
// If a value is missing, fall back to the nearest smaller batch_size.
// Small batch-size variants are dense [2, 4, ..., 48]. Larger long-prefill
// variants are compiled sparsely and dispatch falls back to the nearest
// smaller batch_size when an exact value is not available.
if (entry == nullptr) {
int32_t fallback_batch_size =
specialization.batch_size - kBatchSpecializationStep;
Expand Down
6 changes: 6 additions & 0 deletions xllm/core/kernels/npu/tilelang/fused_gdn_gating_wrapper_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,6 +144,12 @@ TEST_F(TileLangFusedGdnGatingWrapperTest, MatchesTorchReference) {
.num_heads = 32,
.seed = 106,
},
{
.name = "long_prefill_b3533_h32",
.num_batches = 3533,
.num_heads = 32,
.seed = 108,
},
{
.name = "custom_beta2_threshold0p5_b33_h64",
.num_batches = 33,
Expand Down
19 changes: 10 additions & 9 deletions xllm/core/layers/npu_torch/qwen3_gated_delta_net_base.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -393,15 +393,16 @@ torch::Tensor Qwen3GatedDeltaNetBaseImpl::forward(

// Compute gated delta net decay and beta terms.
if (attn_metadata.is_prefill) {
beta = torch::sigmoid(b);
torch::Tensor A_log_exp = A_log_.exp();
torch::Tensor a_float = a.to(torch::kFloat32);
torch::Tensor a_plus_dt = a_float + dt_bias_;
torch::Tensor softplus_out = torch::nn::functional::softplus(
a_plus_dt,
torch::nn::functional::SoftplusFuncOptions().beta(1.0).threshold(20.0));
g = -A_log_exp * softplus_out;
g = g.to(a.dtype()).contiguous();
xllm::kernel::FusedGdnGatingParams gdn_params;
gdn_params.A_log = A_log_;
gdn_params.a = a.contiguous().view({-1, a.size(-1)});
gdn_params.b = b.contiguous().view({-1, b.size(-1)});
Comment on lines +398 to +399
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Apr 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

Prefer using .reshape() over the .contiguous().view() pattern. reshape() is more idiomatic in PyTorch; it returns a view if the tensor is already contiguous and only performs a copy if necessary. This avoids redundant operations and potential memory allocations if the input tensors a and b are already contiguous.

Suggested change
gdn_params.a = a.contiguous().view({-1, a.size(-1)});
gdn_params.b = b.contiguous().view({-1, b.size(-1)});
gdn_params.a = a.reshape({-1, a.size(-1)});
gdn_params.b = b.reshape({-1, b.size(-1)});

gdn_params.dt_bias = dt_bias_;
gdn_params.beta = 1.0f;
gdn_params.threshold = 20.0f;
std::tie(g, beta) = xllm::kernel::fused_gdn_gating(gdn_params);
g = g.squeeze(0).contiguous().view({batch_size, seq_len, a.size(-1)});
beta = beta.squeeze(0).contiguous().view({batch_size, seq_len, b.size(-1)});
Comment on lines +404 to +405
Copy link
Copy Markdown
Contributor

@gemini-code-assist gemini-code-assist Bot Apr 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The calls to .squeeze(0) and .contiguous() are redundant here. Tensors returned by custom kernels are typically contiguous, and view() can handle the reshaping directly from the kernel's output shape (whether it is [total_tokens, hidden] or [1, total_tokens, hidden]) to the target [batch, seq, hidden] shape. Removing these unnecessary calls improves performance in the prefill hot path.

Suggested change
g = g.squeeze(0).contiguous().view({batch_size, seq_len, a.size(-1)});
beta = beta.squeeze(0).contiguous().view({batch_size, seq_len, b.size(-1)});
g = g.view({batch_size, seq_len, a.size(-1)});
beta = beta.view({batch_size, seq_len, b.size(-1)});

} else {
xllm::kernel::FusedGdnGatingParams gdn_params;
gdn_params.A_log = A_log_;
Expand Down
Loading